Generalization and Representational Limits of Graph Neural Networks
This addresses fundamental limitations of GNNs for researchers in graph machine learning, though it is incremental in providing theoretical analysis rather than practical improvements.
The paper proves that certain important graph properties cannot be computed by local-information-based graph neural networks (GNNs), and provides the first data-dependent generalization bounds for message passing GNNs that are much tighter than existing VC-dimension guarantees.
We address two fundamental questions about graph neural networks (GNNs). First, we prove that several important graph properties cannot be computed by GNNs that rely entirely on local information. Such GNNs include the standard message passing models, and more powerful spatial variants that exploit local graph structure (e.g., via relative orientation of messages, or local port ordering) to distinguish neighbors of each node. Our treatment includes a novel graph-theoretic formalism. Second, we provide the first data dependent generalization bounds for message passing GNNs. This analysis explicitly accounts for the local permutation invariance of GNNs. Our bounds are much tighter than existing VC-dimension based guarantees for GNNs, and are comparable to Rademacher bounds for recurrent neural networks.